4 mo.
How fast Uber burned its 2026 AI budget
$1,500
Uber’s new monthly cap, per employee per tool
60T+
Tokens Meta staff used in one 30-day window
$900M
That usage, valued at public API pricing
KEY TAKEAWAYS
• Microsoft canceled most Claude Code licenses in its Experiences and Devices division by June 30, redirecting engineers to GitHub Copilot CLI. The official internal reason is toolchain unification, not cost.
• Uber’s CTO confirmed the company’s entire 2026 AI coding budget was exhausted by April, four months in, after engineers were encouraged to use AI “as much as possible” through an internal leaderboard.
• Since then, Microsoft has set division-level AI token budget targets and Uber has capped spending at $1,500 per employee, per tool, per month, both tracked on internal dashboards.
• Meta ran its own leaderboard, nicknamed “Claudeonomics,” ranking all 85,000 employees by token use. It was pulled within 48 hours of leaking, after one month saw more than 60 trillion tokens consumed.
• Amazon, Adobe, Atlassian, and Citi have each imposed their own throttles this year, according to a 404 Media investigation built on leaked internal documents.
Two Companies, the Same Problem
Microsoft spent six months telling engineers to use Claude Code. Then it canceled most of the licenses. Uber built a leaderboard ranking employees by how much AI they used. Then it capped spending at $1,500 per employee, per tool, per month.
The AI story getting told right now is job replacement. The one actually happening inside these companies is a budgeting problem nobody priced correctly. Microsoft and Uber show the same shape, just at different speeds, and they’re not the only two.
What Microsoft Actually Did
Microsoft opened Claude Code to engineers, PMs, and designers in its Experiences and Devices division, the group behind Windows, Microsoft 365, Outlook, and Teams, in December 2025. Adoption outran Microsoft’s own coding assistant. By June 30, 2026, most licenses were canceled and engineers were redirected to GitHub Copilot CLI.
The official reason, laid out in an internal memo from division head Rajesh Jha, is toolchain unification: Copilot CLI gives Microsoft a product it can shape directly through GitHub rather than relying on a third-party tool. Microsoft has not confirmed cost as a factor. Outlets covering the decision, including The Verge and TechRadar, have noted the timing, coinciding with Microsoft’s fiscal year boundary, and the general cost pressure of token-based billing as reasons to suspect budget played a role, even without an official admission.
Whatever the full reasoning, the budget discipline arrived regardless. Since July, every division has run its own AI token budget target, tracked through an internal usage dashboard, according to an internal memo reported by Storyboard18.
What Uber Actually Did
Uber’s version of the story is more direct, because Uber said the quiet part out loud. CTO Praveen Neppalli Naga told The Information that Uber’s entire 2026 AI coding tools budget, reportedly part of a broader $3.4 billion AI budget for the year, was gone by April, four months in.
Part of the reason was structural. Engineers had been encouraged to use AI “as much as possible,” with usage tracked and ranked competitively on an internal leaderboard. The incentive worked exactly as designed. It just wasn’t designed with a ceiling.
Uber’s fix, reported by Bloomberg in June, was a hard cap: $1,500 per employee, per tool, per month, enforced through an internal dashboard. Uber’s own COO, Andrew Macdonald, later acknowledged on a podcast that it’s “very hard to draw a line” between the AI spending and any resulting improvement to the product.
The Leaderboard Problem
Uber wasn’t alone in gamifying AI use, and the pattern reveals something more specific than overspending. It reveals what happens when a company optimizes for volume instead of value.
Meta ran its own internal leaderboard, reportedly nicknamed “Claudeonomics,” ranking all 85,000 employees by token consumption, complete with badges like “Token Legend” and “Cache Wizard.” In a single 30-day window, staff burned through more than 60 trillion tokens, worth roughly $900 million at public API pricing, with the single top user reportedly averaging 281 billion tokens. Meta pulled the leaderboard within 48 hours of it leaking.
“All motion is not progress, and token usage alone is not a measure of impact of any kind.”
Andrew Bosworth, Meta CTO, in an internal memo following the leaderboard leak
The Pattern Extends Further
Amazon built a comparable internal leaderboard, “KiroRank,” and scrapped it after catching employees gaming their own usage numbers rather than doing more useful work. The details of Amazon, Adobe, Atlassian, and Citi’s responses come from a 404 Media investigation based on leaked Slack messages, internal dashboards, and emails gathered from half a dozen companies.
Adobe ended unlimited Claude access on June 30. Atlassian’s internal AI tooling costs climbed from roughly $5 million to more than $15 million a month over nine months, pushing the company toward tiered usage credits instead of open access. Citi reportedly disabled premium AI models for about a week in late June, though the bank disputed that characterization to 404 Media even after being shown internal documents suggesting otherwise.
The Quote That Explains the Economics
Nvidia’s Bryan Catanzaro, vice president of applied deep learning, put the underlying math plainly in comments to Axios.
His team runs some of the most compute-intensive workloads in the industry, not a typical office function, so the ratio is more extreme than what most companies are dealing with. But the same arithmetic, compute costing more than the people using it, is now showing up in ordinary engineering budgets at companies that never expected to run it.
“For my team, the cost of compute is far beyond the costs of the employees.”
Bryan Catanzaro, Nvidia VP of Applied Deep Learning, to Axios
LUDICARC’S TAKE
None of this is a story about AI eliminating jobs. It’s a story about companies discovering, expensively, that unmetered access to a powerful tool behaves like unmetered cloud computing: usage expands until someone attaches a meter to it. Microsoft, Uber, Meta, Amazon, Adobe, Atlassian, and Citi all ran some version of the same experiment this year, encourage adoption first, worry about cost later, and all seven arrived at some version of the same correction.
What separates the companies handling this well from the ones getting embarrassed by leaked leaderboards isn’t how much they spent. It’s whether they built any way to tell the difference between an employee generating 281 billion tokens of genuine output and one generating 281 billion tokens because the scoreboard rewarded volume. Most of them didn’t, which is why the caps arrived after the spending, not before it.
The real AI story in 2026 isn’t which jobs disappear. It’s which teams figure out how to measure value per token before finance does it for them.
SOURCES
The Information: Uber’s AI Coding Budget, Gone in Months
The Verge: Microsoft cuts Claude Code access, pushes GitHub Copilot CLI
Bloomberg: Uber Caps Employee AI Tool Spending at $1,500 a Month
Fortune: Inside Meta’s “Claudeonomics” leaderboard, and why it got shut down
Axios: Nvidia exec: AI compute costs now dwarf employee costs
404 Media: Companies Are Throttling Employees’ AI Use Because It’s Too Expensive
Storyboard18: Microsoft sets AI token budget targets across divisions
TechRadar: Why Microsoft really cut Claude Code access
